A Guth Labs publication ↗
Back to News

Vercel adds Ampersand for group conversations where silence is an option

Conceptual editorial illustration of a conversation circle beside a quiet AI orb, representing selective participation.
AI-generated editorial illustration for Guth News

Vercel added Humans&’s Ampersand to AI Gateway on November 6, offering free access through November 13. The distribution announcement brings a model trained for group conversations into Vercel’s model-routing service.

This is an additional access channel, not Ampersand’s first release. Humans&’s model card dates version 0.1 to October 20 and describes a text-based collaborator built on GLM 5.3.

Participation becomes a model decision

Ampersand reads messages that are not addressed to it and can choose whether to contribute. Vercel’s integration uses application-defined tools for sending a message or passing until another human message arrives. Developers can call the model as humansand/ampersand-v0.1 through the AI SDK or an OpenAI-compatible interface.

That changes what a successful response looks like. In a group, an accurate answer can still be unwelcome if it interrupts a discussion, repeats what someone has just said or answers a question meant for another person. Choosing not to speak becomes part of the product’s behavior.

Humans& says its training approach uses long-running social simulations to evaluate how a contribution affects individual and shared goals over time. Those simulations are a development method, not evidence that every real group will find the model helpful.

Permissions remain outside the conversation model

The model card warns that Ampersand can misunderstand goals, interrupt, give incorrect information or disclose inappropriate context. It says applications must enforce permissions and validate tool actions, and calls for human confirmation before consequential actions.

The company also reports prompt-injection susceptibility broadly similar to its GLM base, with a successful injection rate above frontier models in the evaluation it used. Social fluency should not be mistaken for a stronger security boundary.

For a team assistant, that distinction is practical. Knowing that one participant wants a file does not establish that the participant is entitled to see it. The application still needs to decide which channels, documents and actions each person can access.

A short trial window

Vercel’s temporary free period creates an opportunity to test conversational timing without treating the promotion as a permanent operating price. Teams can evaluate missed opportunities as well as unnecessary interruptions, and compare behavior in routine exchanges with behavior during disagreements.

Ampersand’s arrival makes a specialized collaboration model easier to reach through an existing gateway. Its value will depend on whether the surrounding product gives it enough context to help while keeping authority and disclosure rules explicit.

Sources and citations

Revision history

Published .